Acoustic Source DOA Tracking Using Deep Learning and MUSIC

Sound source tracking is a classical problem in the array processing field that finds applications from human-robot interaction to acoustic navigation. The emergence of deep learning has produced a new class of state-of-the-art solutions, including the Cross3D, based on a convolutional neural network architecture. In this article, the authors present a new system for sound source tracking via microphone array that modifies the pre-processing stage of Cross3D by using MUSIC maps as inputs, and applying appropriate minor modifications to its architecture. The proposed system, called Spectral Cross 3D, improves overall accuracy by 18%, when compared with the original solution.